<- all tokdocs

Why Vibe-Coded SaaS Breaks at the Second Feature, and What the Data Shows

By , founder of Mavric and Apso

View on TikTok ->

A TikTok on @realcolegordon features a guest labelled "CMO of $30M Company" arguing that a vibe-coded SaaS product breaks once you try to expand it, because the AI made the architecture decisions and nobody understands the code. The core point about architecture holds up, and GitClear's data shows refactoring falling as AI code rose. The clip overstates it with "100% of the time," misuses the word "refactor," and skips the workflow Anthropic documents for keeping architecture decisions with the human.

What the video claims

The guest says the AI coding agent makes the architecture decisions: "Somebody has to," and it will not be the founder. The agent does not know the founder's intent. Then, "what will happen 100% of the time is maybe the initial thing works, maybe it doesn't." When the product needs new users or features, "to move forward, you have to do what's called a refactor," and "you don't have a single engineer that understands your code base." The clip ends mid-sentence.

Does vibe-coded software fail 100% of the time?

The claim qualifies itself. "100% of the time" is followed by "maybe it works, maybe it doesn't," so no outcome is asserted.

The measured record is narrower. The 2025 DORA report, published 23 September 2025 from a survey of nearly 5,000 technology professionals, found 90% of respondents using AI at work. AI adoption had a positive relationship with software delivery throughput and product performance. The negative relationship was with delivery stability: more change failures, more rework and longer time to resolve. That is a real cost, but it is not a universal failure rate.

METR's randomized controlled trial, published 10 July 2025, found 16 experienced open-source developers took 19% longer on 246 real issues when allowed to use AI tools, while estimating afterward that AI had made them 20% faster. METR states that the result does not show AI fails to speed up most developers generally, and its sample was experienced contributors in mature repositories they knew. The clip's scenario is a non-engineer building something new, so the number does not transfer.

Is rebuilding the foundation a refactor?

No. Martin Fowler's definition of refactoring is "a change made to the internal structure of software to make it easier to understand and cheaper to modify without changing its observable behavior." Adding user tiers, adding a feature set or rebuilding a data model for multi-tenancy changes observable behavior. What the clip describes is a re-architecture or rewrite.

The difference affects cost. A refactor is bounded and can be checked against a test suite. A rewrite has no fixed scope and no oracle, and calling it a refactor makes it sound cheaper than it is.

Does AI-generated code reduce refactoring?

Yes, according to GitClear's AI code quality research, which covers 211 million changed lines from January 2020 through December 2024. Refactored ("moved") lines fell from 24.1% of changed lines in 2020 to 9.5% in 2024. Copy-pasted lines rose from 8.3% to 12.3%, a 48% relative increase. 2024 was the first year on record where newly introduced duplicate code exceeded refactoring activity.

DORA's 2025 report supports the architecture argument. Teams in loosely coupled architectures with fast feedback loops see gains from AI, while teams in tightly coupled systems with slow processes see little or no benefit. A vibe-coded first version is unlikely to have loose coupling or a fast feedback loop, because nobody chose either one.

Do developers understand AI-written code?

The 2025 Stack Overflow Developer Survey points to a comprehension gap even among professionals:

  • 84% use or plan to use AI tools, up from 76% the prior year.
  • 3.1% highly trust the accuracy of AI output, and 45.7% actively distrust it.
  • 66% name "AI solutions that are almost right, but not quite" as their top frustration.
  • 45.2% say debugging AI-generated code takes more time than writing it.

Those developers can read the code. A founder who cannot has the same gap and no way to detect it.

How to keep architecture decisions with the human

The clip treats the AI making architecture decisions as inevitable. Anthropic's Claude Code best practices describe a different default. The documented workflow has four phases: explore, plan, implement and commit. Plan mode is a separate permission state where the agent reads files and answers questions without writing anything, and the page recommends opening the plan in an editor and changing it before approving.

The same page says to record "architectural decisions specific to your project" in a CLAUDE.md file that loads at the start of every session. It also recommends having the agent interview you and write a spec to SPEC.md before implementation. It names the failure the clip describes, "the trust-then-verify gap," and gives the rule "if you can't verify it, don't ship it."

Who is the speaker?

The caption calls him "CMO of $30M Company" and names neither the person nor the company. Cole Gordon, who owns the account, is CEO of Closers.io, so the arrow points at an unnamed guest. I could not find a primary source for the $30M figure.

"Vibe coding" itself is recent. Andrej Karpathy coined the term in February 2025, and Collins English Dictionary named it Word of the Year on 6 November 2025.

Bottom line

The architecture argument is the strongest part of the clip, and GitClear and DORA both back it. "100% of the time" is not a measured claim, and the fix the clip calls a refactor is a rewrite. Claude Code's documented plan, spec and CLAUDE.md workflow is the direct answer to an agent making architecture decisions nobody chose.

Sources

Published October 2, 2026. Based on a TikTok by @realcolegordon. Full video notes (markdown)